[KongchangAI]
Modeld1-omni / Liquid d1-omni-600M

d1-omni-600M

An experimental multimodal decision model released by Liquid AI, built on a bidirectional encoder architecture. It processes text, image and audio inputs but does not generate text, instead directly outputting a decision distribution. The model has 587M parameters.

Timeline (last 90 days)

Oct 14

LiquidAI's d1-omni-600M runs on CPU, and all 18 newly added blocks were false positives, with recall remaining at the upstream 0.319.

Unverified50%
Oct 13

The d1-omni-600M model has about 587 million parameters, is based on a small bidirectional encoder with added vision and audio encoders, supports text plus images or text plus up to 30 seconds of audio, and Liquid calls it an early research version.

Unverified50%
Oct 13

Liquid AI's d1-omni-600M is explicitly an early research version, and no latency data has been published.

Unverified50%
Oct 12

d1-omni-600M 包含 381M 共享主干加决策头、94M SigLIP2 视觉编码器(来自 LFM2.5-VL-450M)和 112M 17 层 FastConformer 音频编码器。

Unverified50%
Oct 12

d1-omni-600M 基于 LFM2.5-Encoder-350M 双向编码器构建,而非在因果解码器上接分类头,上下文窗口为 16,384 tokens。

Unverified50%
Oct 12

d1-omni-600M 支持三种问题类型:noul(是/否/中立二元判断)、choice(从命名选项中选一个)和 score(从 2 到 10 个有序等级中选一个),同一状态下所有问题共用一次前向传播。

Unverified50%
Oct 12

d1-omni-600M 的标签支持 15 种语言,包括英语、德语、阿拉伯语、印地语、日语和中文。

Unverified50%
Oct 12

d1-omni-600M 图像输入时,状态和问题文本会被截断到 896 tokens;音频输入限制为单条 16 kHz 单声道片段、最长 30 秒,且音频训练数据仅来自英语使用者与助手之间的请求;图像和音频不能在同一次请求中同时携带,否则抛出 ValueError。

Unverified50%
Oct 12

d1-omni-600M 在 Civil Comments 内容审核基准上得分为 95.8,为该对比表中最高。

Unverified50%
Oct 12

Liquid 官方表示不报告 d1-omni-600M 的推理速度数据,博客中提到的亚 50ms 延迟仅适用于 d1-3B。

Unverified50%

7 more timeline events

All Facts (17)

Unverified

LiquidAI's d1-omni-600M runs on CPU, and all 18 newly added blocks were false positives, with recall remaining at the upstream 0.319.

50%
Unverified

Liquid AI's d1-omni-600M is explicitly an early research version, and no latency data has been published.

50%
Unverified

The d1-omni-600M model has about 587 million parameters, is based on a small bidirectional encoder with added vision and audio encoders, supports text plus images or text plus up to 30 seconds of audio, and Liquid calls it an early research version.

50%
Unverified

d1-omni-600M 基于 LFM2.5-Encoder-350M 双向编码器构建,而非在因果解码器上接分类头,上下文窗口为 16,384 tokens。

50%
Unverified

d1-omni-600M 支持三种问题类型:noul(是/否/中立二元判断)、choice(从命名选项中选一个)和 score(从 2 到 10 个有序等级中选一个),同一状态下所有问题共用一次前向传播。

50%
Unverified

d1-omni-600M 的标签支持 15 种语言,包括英语、德语、阿拉伯语、印地语、日语和中文。

50%
Unverified

d1-omni-600M 图像输入时,状态和问题文本会被截断到 896 tokens;音频输入限制为单条 16 kHz 单声道片段、最长 30 秒,且音频训练数据仅来自英语使用者与助手之间的请求;图像和音频不能在同一次请求中同时携带,否则抛出 ValueError。

50%
Unverified

d1-omni-600M 在 Civil Comments 内容审核基准上得分为 95.8,为该对比表中最高。

50%
Unverified

Liquid 官方表示不报告 d1-omni-600M 的推理速度数据,博客中提到的亚 50ms 延迟仅适用于 d1-3B。

50%
Unverified

d1-omni-600M 各量化格式的磁盘占用:原始 F32 safetensors 2.35 GB,GGUF F16 约 1.2 GB,GGUF Q8_0 约 670 MB,ONNX q8(纯文本)0.41 GB,ONNX q4(纯文本)0.24 GB,ONNX 语音栈 0.67 GB。

50%
Unverified

ONNX 维护者实测发现 4-bit 决策权重会使语音答案概率偏移最多 0.52,且 Liquid 模型卡建议 GPU 上避免 bfloat16,因其会在 0.8% 文本和 1.7% 音频行上翻转首选答案。

50%
Unverified

对于 d1-omni-600M,建议在树莓派等边缘设备上使用 Q8_0 GGUF 量化,避免低量化以防止首选答案翻转,并为激活值预留 20%~30% 的内存余量。

50%
Unverified

在 Liquid AI 的 Decision Index 0.2.1 上,d1-omni-600M 得分 15.95,d1-3B 得分 48.57,Decider 2B 得分 28.97。

50%
Unverified

Liquid AI d1-omni-600M 参数规模为 587M,Q8_0 GGUF 量化后约 670 MB,llama.cpp 已合并支持,可在树莓派 5 上通过 llama-server 命令运行。

50%
Unverified

d1-omni-600M 包含 381M 共享主干加决策头、94M SigLIP2 视觉编码器(来自 LFM2.5-VL-450M)和 112M 17 层 FastConformer 音频编码器。

50%
Unverified

Liquid AI's Open d1 series includes d1-3B (text + image) and d1-omni-600M (text + image or text + audio, experimental).

50%
Unverified

Liquid AI 于 10 月 7 日发布 d1-omni-600M 与更大的 d1-3B,官方将其定位为实验性检查点。

50%

Source Articles

AI Decision Models: When Your AI Should Return Probabilities, Not ParagraphsOct 13
Liquid d1-omni-600M:树莓派上的多模态决策模型实战Oct 12